Can AI help us predict extreme weather?

AI models are starting to revolutionize weather forecasting.

Subscribe to our channel and turn on notifications (🔔) so you don’t miss any videos:

This video is sponsored by Microsoft Copilot for Microsoft 365. Microsoft has no editorial influence on our videos, but their support makes videos like these possible.

We’ve learned how to predict weather over the past century by understanding the science that governs Earth’s atmosphere and harnessing enough computing power to generate global forecasts. But in just the past three years, AI models from companies like Google, Huawei, and Nvidia that use historical weather data have been releasing forecasts rivaling those created through traditional forecasting methods.

This video explains the promise and challenges of these new models built on artificial intelligence rather than numerical forecasting, particularly as it relates to the ability to foresee extreme weather.

Here are the papers that describe the models mentioned in the video.

Google’s GraphCast:
Huawei’s Pangu-Weather:
Nvidia’s FourCastNet:

Here is the announcement of the ERA5 dataset, released by the European Centre for Medium-Range Weather Forecasts in 2020:

We interviewed Dr. Aaron Hill over email for this video. Hill is involved in developing responsible AI for environmental science via AI2ES:

Google has also developed a weather forecasting model called Nowcasting, which is already embedded in its weather products specifically for short-term precipitation forecasts:

If you’re interested in learning more about the history of how we developed weather forecasting, I’d recommend The Weather Machine by Andrew Blum:

Subscribe to our channel! is a news website that helps you cut through the noise and understand what’s really driving the events in the headlines. Check out

Watch our full video catalog:
Follow Vox on Facebook:
Or Twitter:


#extreme #extremeweather #lightning #weather #weatherwtf

What do you think?

336 Points

Written by weatherwtf


  1. A tool, perhaps that can improve our understanding as to why modelling physics sometimes gets the probabilities wrong. A way to drill into the data prediction patterns that may have, until now missed some interactions.

  2. Vox delivering another excellent bit of content. Been loving these AI-related videos which have been shared over the last few months. Super interesting!

  3. My repsonse to the title of the video: lets assume it did predict at 99.99% accuracy, what action can we take on that prediction? For that matter do we even need ai to predict our weather?

  4. So here's what you do to make the perfect weather A.I.:
    You need 3 of them.
    One model is the safe model that they have, it focuses only on the safest and most correct calculations. The second model needs to be an AI that's specifically tailored to predict bizarre weather events and storm intensity.
    Finally, the third model: Is a copy of the first model, but every time the second model makes a correct prediction, its then fed into the third model's data and explained exactly why that happened. Eventually, the third model will be able to accurately predict both normal and extreme weather events with terrifying precision.

  5. Didnt get how AI can produce more ensemble forecasting and has less time constraints compared to traditional forecasting. If AI does not run the actual physics simulation, how are more ensembles more useful? It generates additonal ensembles based on past data and probability?

  6. I feel it’s important not to leave out the physical calculations involved in weather forecasting. For example, try throwing a ball with certain conditions and predict where it will land. For an AI to properly learn and predict, it will need hundreds if not thousands of examples, but just doing the math will get the answer in a jiffy. You can learn from experience and quickly predict the results but you will never be able do so with certainty and also accurately predict exotic events.
    The video does makes a good point saying that AI would have trouble predicting exotic events. Say most of the time the AI was trained using a dataset where most of the ball throws landed 20 yards away, but only once or twice did it land 60 yards away. AI would be so accustomed to those 20 yard throws that it almost disregards the exotic outcomes. And when that does actually happen, the AI will just say it landed 20 yards away. Where as just calculating the results will get you an actual good prediction.
    I’m not saying AI should never be used in weather forecasting, in fact it’s great at doing so. What I’m trying trying to say is that it shouldn’t be exclusively used to do so without the physics and math, rather we should use them in combination. I can see new technologies will ever bring us far more accurate weather forecasts and hope that we will get to use them soon.

  7. AI told me to join Extinction rebellion to try save human race ! after querying latest climate data and its consequences ,that's an improvement on their previous `we can still hit 1,5 limit bollux .

  8. The thing is, as a forecaster for my YouTube channel, of course it’s hard and I do believe there is some chance that AI can get it with enough practice however AI will never be able to have the sequential and the ability to determine complex weather patterns like people can right now. Plus, I do want to note how forecasting is so hard with weather models. I’m positive you were referring to the GFS models when saying the update every 6 hours which is true but they are low quality and a lot of people kinda try and stray away from GFS but there are high quality models that update every hour like the HRRR. GFS is so inaccurate as far out as 3 days out because of the fact that they are trying to forecast long distance weather patterns far ahead of time and because conditions can change in an instant, that’s even more amplified hundreds of hours out on a model so with each new model run, things are bound to be different, especially with GFS. As a result, GFS is never used for in depth forecasting. It’s important to note how these models change because a human with the experience and reasoning for weather forecasting that AI will likely almost never have. I don’t doubt that there will likely be AI models in the future but with what I have seen with the weather and how much can change, I doubt they will be accurate without human input. I really hope this made sense and the grammar isn’t bad cause I was writing this while half asleep

  9. Thank God for these brilliant minds all throughout our history ,the ones we have now and to the future ones that are coming. If not for them the convenience they provide to all of us won't be the same. Thank you 👏

  10. 6:11 the graph here is AWFUL. What is an extreme weather event? The only way I can see this graph being correct is if you're considering the cost of damages to distinguish between extreme and non extreme weather events. Then you gotta consider inflation AND all the humans that decided to build places in hurricane country after AC was invented.
    Also, the graph starts at 1970 despite having good records back to the mid 19th century, this is a common trick.
    Can we get a proper definition of extreme weather events?

Tornadoes, severe weather hit northwest Ohio Thursday | WTOL 11 Team Coverage – 3/14, 11 p.m.